These data present a prime facie challenge to standard structural assumptions: the fact that APs and possessives intervene between the head noun and its complements is not compatible with either of the X' schemata [ spec [ head comps ]] and [[ head comps ] spec ]. The most well-established analysis of NP structure in Celtic maintains standard X' theoretic assumptions by positing head movement of the N to a dominating functional projection Num in all NPs [Rouveret 1990][Rouveret 1994]: some analyses involve further raising of N to Agr or D in possessive constructions [Duffield 1996]. The structural syntax shown schematically in (8) accounts straightforwardly for N (AP*) (Poss) (Comps) word order.
Raising N over AP accounts for the occurrence of adjectival modifiers between a head and complements/specifiers, as in (4) and (7). Likewise the raising of N to a position outside and dominating NP accommodates the observed head - specifier - complement order which is otherwise problematic. With head movement, word order within the extended nominal projection is derived in a fashion precisely similar to that of VSO clausal order (which involves raising V to I). In sum, analyses in this style postulate at least two (D, Num) and sometimes three (D, Num, Agr) functional projections within nominal structure, with obligatory N raising forced by some set of assumptions about affixation or the strength/weakness of abstract features.
The head raising account can be expressed in the monostratal constraint based formalism of LFG despite the absence of derivational mechanisms. [Bresnan 1997a] adopts a rather rich view of constituent structure in endocentric configurational languages. The theory posits both lexical and functional categories, but within a strongly lexicalist perspective and provides a clearly articulated and highly constrained view of the role of functional categories and of the relationship between lexical projections and their extended (functional) projections.
The canonical mapping from nodes to f-structures in configurational endocentric structures is expressed by a set of universal principles of (unmarked) structure-function association making reference to sets of grammatical functions including adjuncts, the non-discourse argument functions (CF) and the discourse functions (SUBJ, TOPIC, FOCUS) [Bresnan 1997a].
The central insight of head movement analyses is modelled in this theory by the extended head theory (for this formulation see [Bresnan 1997b])3
The notion of extended head permits us to model the head movement account of Welsh NPs, capturing the key aspects of surface phrase structure directly. If nouns are actually functional Num heads and APs adjoin to NP, then we predict that APs are postnominal and precede complements and specifiers of NP. We also predict that possessors (specifiers in constituent structure) intervene linearly between head noun and complements. Below we illustrate with (2) and (4), marking co-heads ^=v. The unmarked or canonical function for the specifier of a lexical category is that of adjunct. In Welsh (and the other Celtic languages), the Spec,NP position is associated with the possessive.4 I will assume without further comment that POSS is a SUBJective function and thus that these languages depart from the unmarked mapping for specifier of NP.5
The head movement accounts of Celtic noun phrase structure suffer from a number of serious technical difficulties, and are empirically inadequate in various respects when a full range of data is considered. Space precludes providing a detailed discussion of this, but a number of problems with the accounts will emerge as we discuss the alternative in the following sections. The fully lexicalised version in fact avoids a number (but not all) of these empirical problems: for example, no difficulty arises in ensuring that head movement is obligatory in the syntax if nouns are lexically specified as members of the functional category Num! However a major question arises (in each formalisation) concerning the status of this second functional category within nominal structures.
The most well-established `head-raising' analyses (of clausal structure) involve two clear, empirical arguments. Firstly, the functional projection itself can be distributionally motivated as the structural position for certain closed class, functional elements, and secondly, a strong argument can be made for an inflectionally defined class of substantive (lexical) heads appearing in a functional position on clear morpho-syntactic grounds - a typical example would be a case in which finite verbs appear in I or C while non-finite verbs appear in V. The difficulty, however, is that there is no apparent morphosyntactic basis for head raising in Welsh NPs. Head movement is obligatory in noun phrases, irrespective of the presence of determiners (so movement cannot be argued to follow from the need to host affixal determiners), and independent of number marking on the noun (so movement cannot be argued to follow from the need to host a number affix). Neither can N raising be related to the occurrence of a possessive (e.g. the requirement to stand in some structural relationship (perhaps c-command) to the possessive), since it is independent of the occurrence of possessives. Head movement accounts of Celtic NPs variously appeal to abstract affixes and features or the strength/weakness of purely abstract morphological features (without external realisation) to induce head movement as required. Similarly, in well motivated extended head analyses in LFG, the functional categories typically correspond to sets of closed class elements (expressing collections of morphosyntactic features) and/or inflectionally defined subclasses of lexical or open class categories (so typically, there is variation in the position of verbal elements, and so on).
The primary motivation for treating nouns as heads of a Num projection, in either the derivational head movement account, or in the lexicalised extended head analysis, comes only from the observed word order facts: the analyses permit a certain standard set of assumptions about constituent structure to be maintained in the light of empirical difficulties. I conclude that the analysis is therefore undermotivated, and pursue an alternative hypothesis in the rest of this paper.6